NVIDIA has released two open source projects designed to run managed Kubernetes on GPU infrastructure: NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR). Recently announced as components of the NVIDIA DSX AI Factory platform, these two software packages are distributed under the Apache 2.0 license and are in public preview.
Kubernetes is a tool that organizes containerized applications across multiple servers. NVKE constitutes the control plane for a managed Kubernetes service. Installed by an ‘AI factory’ on its own graphics processing unit (GPU) infrastructure, it exposes a REST API and controllers on a management cluster. Clients receive standard clusters, which they use with the kubectl tool, while the operator administers them via the NVKE API or the nvkectl command line.
The project repository describes NVKE as ‘a self-hosting factory for Kubernetes clusters’, not as a distribution. The software is based on upstream Kubernetes, on kubeadm and on Cluster API, and is written in Go. It does not depend at runtime on NVKR.
NVKR, for its part, is a compilation system based on declarative definitions. It produces executables, container images, Helm charts, and node images for amd64 and arm64 architectures. Its ‘recipes’ define validated combinations of software versions and GPU generations. Hosted workflows associate signatures, SBOMs, and provenance information, linked to OCI references pinned by digest. NVKR can also generate artifacts for other Kubernetes management systems. The repository does not yet mention any qualified version.
A five-step deployment
According to the technical post published by NVIDIA, installation goes through five steps. The nvkectl bootstrap site command first launches a temporary single-node kind cluster on a bootstrap host. NVKE then requests three servers from NICo, which starts them from the node image, after which the temporary cluster is removed. NICo manages the bare metal hardware lifecycle, and CAPNICo, a Cluster API provider, links NICo to NVKE. Operators then publish a machine image, a cluster version, and a node pool version, which clients select by name.
Each client cluster has its own control plane, its own etcd database, and its own node pools. NVKE installs OVN-Kubernetes and NodeLocal DNSCache on them. A node must present an attested SPIFFE identity before obtaining its connection credentials, and the README indicates that SPIRE relies on a TPM for this attestation. etcd encryption is available per cluster, and an OpenTelemetry-based Helm chart exports logs, metrics, and traces via OTLP. Three complementary projects integrate with client clusters: AI Cluster Runtime (AICR) for the GPU environment, NodeWright for node packages, and NVSentinel for automated repairs.
Certification and preview limitations
NVKE is presented as Certified Kubernetes AI Platform and Certified Kubernetes Installer, with conformance tests on Kubernetes 1.36. No throughput or provisioning time benchmarks are published. Once installation is complete, the NVKE API responds in plain HTTP on port 30080, with authentication disabled, and no procedure for updating or backing up the management plane is documented.
Several players are already integrating the two projects. NVIDIA’s cloud partners Firmus and SK Telecom integrate them into their AI factories, while Mirantis, Rafay, Spectro Cloud, OpenNebula Systems, and vCluster add components to their managed Kubernetes offerings. Netris enables NVKE for its network automation. A local evaluation requires at minimum 6 CPUs, 12 GiB of free memory, 100 GiB of disk space, Docker, and support for a virtual TPM, without a GPU or an NVIDIA account. A real installation requires a NICo deployment with registered capacities, a build host with KVM, an HTTP server, and a container registry.
